It^ocalculus in a nutshell Vlad Gheorghiu Department of Physics Carnegie Mellon University Pittsburgh, PA 15213, U.S.A. April 7, 2011 Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 1 / 23 Outline 1 Elementary random processes 2 Stochastic calculus 3 Functions of stochastic variables and It^o'sLemma 4 Example: The stock market 5 Derivatives. The Black-Scholes equation and its validity. 6 References A summary of this talk is available online at http://quantum.phys.cmu.edu/QIP Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 2 / 23 Elementary random processes Elementary random processes Consider a coin-tossing experiment. Head: you win $1, tail: you give me $1. Let Ri be the outcome of the i-th toss, Ri = +1 or Ri = −1 both with probability 1=2. Ri is a random variable. 2 E[Ri ] = 0, E[Ri ] = 1, E[Ri Rj ] = 0. No memory! Same as a fair die, a balanced roulette wheel, but not blackjack! Pi Now let Si = j=1 Rj be the total amount of money you have up to and including the i-th toss. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 3 / 23 Elementary random processes Random walks This is an example of a random walk. Figure: The outcome of a coin-tossing experiment. From PWQF. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 4 / 23 Elementary random processes If we now calculate expectations of Si it does matter what information we have. 2 2 E[Si ] = 0 and E[Si ] = E[R1 + 2R1R2 + :::] = i. The random walk has no memory beyond where it is now. This is the Markov property. The random walk has also the martingale property: E[Si jSj ; j < i] = Sj . That is, the conditional expectation of your winnings at any time in the future is just the amount you already hold. Pi 2 Quadratic variation: j=1(Sj − Sj−1) . You either win or lose $1 after each toss, so jSj − Sj−1j = 1. Hence the quadratic variation is always i. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 5 / 23 Elementary random processes Brownian motion Now change the rules of the game: allow n tosses in a time t. Second, the size of the bet will not be $1 but $pt=n. Again the Markov and martingale properties are retained and the 2 Pn 2 q t quadratic variation is still j=1(Sj − Sj−1) = n n = t. In the limit n ! 1 the resulting random walk stays finite. It has an expectation, conditioned on a starting value of zero, of E[S(t)] = 0, and a variance E[S(t)2] = t. The limiting process as the time step goes to zero is called Brownian motion, and from now on will be denoted by X (t). Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 6 / 23 Elementary random processes Figure: A series of coin-tossing experiments, the limit of which is a Brownian motion. From PWQF. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 7 / 23 Elementary random processes Most important properties Continuity: The paths are continuous, there are no discontinuities. Brownian motion is the continuous-time limit of our discrete time random walk. Markov: The conditional distribution of X (t) given information up until τ < t depends only on X (τ). Martingale: Given information up until τ < t the conditional expectation of X (t) is X (τ). Quadratic variation: If we divide up the time 0 to t in a partition with n + 1 partition points tj = jt=n then n X 22 X (tj ) − X (tj−1) ! t: (Technically \almost surely.") j=1 Normality: Over finite time increments tj−1 to tj , X (tj ) − X (tj−1) is Normally distributed with mean zero and variance tj − tj−1. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 8 / 23 Stochastic calculus Stochastic integral Let's define the stochastic integral of f with respect to the Browinan motion X by n Z t X W (t) = f (τ)dX (τ) := lim f (tj−1)(X (tj ) − X (tj−1)) ; n!1 0 j=1 jt with tj = n . The function f (t) which is integrated is evaluated in the summation at the left-hand point tj−1, i.e. the integration is non anticipatory.. This choice of integration is natural in finance, ensuring that we use no information about the future in our current actions. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 9 / 23 Stochastic calculus Stochastic differential equations Stochastic differential equations: The shorthand for a stochastic integral comes from “differentiating” it, i.e. dW = f (t)dX : For now think of dX as being an increment in X , i.e. a Normal random variable with mean zero and standard deviation dt1=2. Moving forward, imagine what might be meant by dW = g(t)dt + f (t)dX ? It is simply a shorthand for Z t Z t W (t) = g(τ)dτ + f (τ)dX (τ): 0 0 Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 10 / 23 Stochastic calculus The mean square limit 2 Pn 2 Examine the quantity E j=1 (X (tj ) − X (tj−1)) − t , where tj = jt=n. Because X (tj ) − X (tj−1) is Normally distributed with mean zero and 2 variance t=n, i.e. E (X (tj ) − X (tj−1)) = t=n, one can then easily 1 show that the above expectation behaves like O( n ). As n ! 1 this tends to zero. We therefore say n X 2 (X (tj ) − X (tj−1)) = t j=1 in the \mean square limit". This is often written, for obvious reasons, as Z t (dX )2 = t: 0 Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 11 / 23 Functions of stochastic variables and It^o'sLemma Functions of stochastic variables If F = X 2 is it true that dF = 2XdX ? NO! The ordinary rules of calculus do not generally hold in a stochastic environment. Then what are the rules of calculus? Figure: A realization of a Brownian motion and its square. From PWQF. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 12 / 23 Functions of stochastic variables and It^o'sLemma It^o'sLemma: A physicist's derivation. Let F(X) be an arbitrary function, where X (t) is a Brownian motion. Introduce a very, very small time scale h = δt=n so that F (X (t + h)) can be approximated by a Taylor series: F (X (t + h)) − F (X (t)) = dF 1 d2F (X (t + h) − X (t)) (X (t)) + (X (t + h) − X (t))2 (X (t)) + :::: dX 2 dX 2 From this it follows that (F (X (t + h)) − F (X (t))) + (F (X (t + 2h)) − F (X (t + h))) + ::: (F (X (t + nh)) − F (X (t + (n − 1)h))) = n X dF = (X (t + jh) − X (t + (j − 1)h)) (X (t + (j − 1)h)) dX j=1 n 1 d2F X + (X (t)) (X (t + jh) − X (t + (j − 1)h))2 + ::: 2 dX 2 j=1 Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 13 / 23 Functions of stochastic variables and It^o'sLemma d2F d2F We have used the approximation dX 2 (X (t + (j − 1)h)) = dX 2 (X (t)), consistent with the order we require. The first line becomes simply F (X (t + nh)) − F (X (t)) = F (X (t + δt)) − F (X (t)). R t+δt dF The second is just the definition of t dX dX . 1 d2F Finally the last is 2 dX 2 (X (t))δt in the mean square sense. Thus we have F (X (t + δt)) − F (X (t)) = Z t+δt dF 1 Z t+δt d2F (X (τ))dX (τ) + 2 (X (τ))dτ: t dX 2 t dX Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 14 / 23 Functions of stochastic variables and It^o'sLemma Can extend this over longer timescales, from zero up to t, over which F does vary substantially, to get Z t dF 1 Z t d2F F (X (t)) = F (X (0)) + (X (τ))dX (τ) + 2 (X (τ))dτ: 0 dX 2 0 dX This is the integral version of It^o'sLemma, which is usually written in differential form as dF 1 d2F dF = dX + dt: dX 2 dX 2 Do a naive Taylor series expansion of F , disregarding the nature of X : dF 1 d2F F (X + dX ) = F (X ) + dX + dX 2: dX 2 dX 2 To get It^o'sLemma, consider that F (X + dX ) − F (X ) was just the 2 R t 2 \change in" F and replace dX by dt, remembering 0 (dX ) = t. This is NOT AT ALL rigorous, but has a nice intuitive feeling. Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 15 / 23 Functions of stochastic variables and It^o'sLemma Coming back to F = X 2 and applying It^o'sLemma, we see that F satisfies the stochastic differential equation dF = 2XdX + dt: In integrated form Z t Z t Z t X 2 = F (X ) = F (0) + 2XdX + 1dτ = 2XdX + t 0 0 0 . Therefore Z t 1 1 XdX = X 2 − t: 0 2 2 Vlad Gheorghiu (CMU) It^ocalculus in a nutshell April 7, 2011 16 / 23 Example: The stock market The lognormal random walk A stock S is usually modelled as dS = µSdt + σSdX ; where µ is called the drift and σ the volatility.